A Novel Approach for Spatially Controllable High-Frequency Forecasts of Park Visitation Integrating Attention-Based Deep Learning Methods and Location-Based Services

Abstract

Forecasting tourism volume can provide helpful information support for decision-making in managing tourist attractions. However, existing studies have focused on the long-term and large-scale prediction and scarcely considered high-frequency and micro-scale ones. In addition, the current approaches are limited regarding forecasting the visitor volume of a designated sub-area in a tourist attraction. This sub-area forecast can assist local-scaled managing decisions of tourist attractions, particularly for large-scale tourist attractions. Therefore, to achieve high-frequency forecasts of tourist volume for finer scale areas such as parks and their sub-areas and generate more controllable and flexible forecasts, this study developed a novel method that incorporates a forecasting model composed of multiple deep learning components and a designed control mechanism. The control mechanism produces high-temporal-resolution sequences of tourist volume for designated sub-areas, and the forecasting model is built on an attention-based deep-bidirectional neural network to better capture the long-range dependencies of the sequence and enhance the forecasting accuracy and robustness. The experimental research was performed at Taiyangdao Park and its two designated sub-areas to validate the effectiveness and superiority of the proposed method compared to other widely used deep-learning methods; three types of performance evaluations were adopted including fitting methods, error measures, and Diebold–Mariano tests. The results demonstrated that the proposed method provided outstanding performance in high-frequency forecasts and yielded more desired forecasting outcomes than other widely used forecasting methods. Furthermore, the comparison with the performances of various other deep learning models provide insights concerning their forecasting capacity; for instance, bidirectional RNN models tend to achieve better forecasts than general RNN models in the high-frequency forecasts. The proposed method has significant practical applicability in aiding short-term micro-scale management decisions and can also serve as an alternative approach in the field of tourist volume forecasting.

1. Introduction

Tourist volume forecasting is an emerging research area in tourism management and tourism geography. With the development and application of AI algorithms and techniques in tourism volume forecasting, forecast accuracy has been significantly enhanced. Meanwhile, its practical implications in informing tourism-related decision-making have gained significant recognition by many researchers and practitioners [1]. Concretely, the practical value of accurate tourism volume forecasting for tourism attractions is embodied in aiding three aspects of tourism-related decisions, which are management decisions, e.g., crowd control, staffing, and scheduling arrangement in the tourist attraction [2,3]; economic decisions, e.g., formulating pricing strategies [3,4]; planning decisions, e.g., subsequent infrastructure planning for the tourist attraction [5]. The temporal characteristics can be highly differentiated according to the type of decision. For instance, planning decisions such as infrastructure plans can last for years, in contrast, management decisions about crowd control and staffing arrangement usually occur intra-day. Therefore, to better support related decision-making, the temporal characteristics of tourist volume forecast should be in accordance with the decision to be made.

Tourism volume forecasts can be divided into long-term, medium-term, and short-term from the perspective of the forecasting horizon. Long-term forecasts have a relatively long-time span of prediction that can be equal to or more than a year [6,7,8]; medium-term forecasts have a shorter forecasting horizon compared to the long-term, approximately six-months or a quarter [9,10]; the horizon of short-term forecasts can be as short as one month [11,12]. In terms of temporal resolution, the most common resolution for tourism volume forecasting is monthly [8,13], weekly [14,15,16,17], and daily [9,11,12].

Existing studies of tourist volume forecasts focused on long- or medium-term predictions, capturing the long-running trend of the tourist volume variation. These studies are crucial for understanding the long-term tourist behavior and supporting strategic decision-making. However, such approaches are not competent for extracting high-temporal-resolution information over a short period to advise short-run decision-making. Although a few researchers have realized the importance of short-term forecasts, and studies on such predictions have emerged [18,19], short-term forecasts, particularly the ones that can reflect intraday fluctuations, are still considerably scarce in the field of tourism volume forecasts. The importance of understanding the time-of-day fluctuation of tourist volume has been emphasized by UNESCO World Heritage Sustainable Tourism Toolkit (http://whc.unesco.org/sustainabletourismtoolkit (accessed on 4 May 2021)), which holds that the information about such fluctuations can help enhance the quality of tourist experience and mitigate pressure, etc. Therefore, the temporal resolution of tourist volume forecasting needs to be improved to guide some short-term management operations. In particular, the prediction time granularity must be increased to a higher level than daily to reflect the intraday variation of tourist volume. In this study, tourist volume forecasts with frequencies that are finer than hourly are denoted as high-frequency (or high-temporal-resolution) forecasts. Such forecasts can provide more details regarding upcoming visits to a tourist attraction, i.e., the variation of tourist volume during a minimal period, and allow the forecasting horizon to be as short as a single day, providing crucial information support to advise short-term decision making.

Spatial scale is another critical aspect related to tourist volume forecasting, where forecasts of a more specific area at a micro scale can provide more valuable information in managerial decisions such as crowd control. However, most related studies have focused on forecasting at a national [20], provincial [21], and urban scale [13]. While studies on site scale are mostly focused on large-sized tourist attractions such as Jiuzhaigou (650.7 km²) [10] and Huangshan Mountain Area (160.6 km²) [11]. These forecasts for the entire tourist attraction can certainly contribute to estimating the visitation variation of the entire site. However, actual managerial decisions such as crowd control are more dependent on specific forecasting information of finer local areas, particularly, the forecasting of crowd-prone sub-areas, which can provide well-targeted support to local crowd management.

In general, most related studies rarely consider high-frequency and small-scale forecasts, and the existing approaches are not competent for achieving such predictions. In particular, forecasting for a micro-scale area such as a sub-park area can provide valuable information for local management decisions such as local crowd control, accident prevention, and staff arrangement. Therefore, to address this gap, we proposed a novel method for such forecasts that incorporates a proposed control mechanism for the forecasting scope and a forecasting model built of multiple deep-learning components. We designed the control mechanism for actualizing tourist volume forecasts of designated sub-area in the park making the forecast more controllable and flexible. The forecasting model receives a high-frequency sequence of tourist volume generated from the control mechanism and completes the forecasting. The model comprises an attention mechanism and a deep bidirectional GRU to identify relevant features of high-temporal-resolution tourist volumes and improve the model’s capability of learning long-term dependency in the tourist volume sequence.

The remainder of this article is structured as follows. Section 2 (literature review) reviews the related works. Section 3 (materials and methods) introduces the proposed control mechanism and forecasting model for high-frequency forecast of tourist volume. Section 4 (experiment) introduces the experimental procedures, study areas, data preparation, model setting, and performance evaluations. Section 5 (results and discussion) analyzes and discusses the experimental results. Finally, the conclusion section summarizes the study and discusses its contributions and practical implementations.

2. Literature Review

2.1. Approaches for Tourist Volume Forecasting

Tourism volume forecasting methods are commonly divided into time-series, econometrics, and artificial intelligence models [22]. The time-series models generate predictions based on past values (e.g., ARIMA and its variants, etc.,) and the econometrics analysis models explore the causal relationship between tourist volume and relevant factors (e.g., ADL, VAR, TVP and PDR, etc.,). As of now, the time-series and econometrics models are dominantly used in the field of tourist volume forecasting [23], and many of the forecasts achieved desired outcomes [24,25,26,27,28,29]. However, a limitation of these models is that they are sensitive to the stability of past time-series patterns and economic structures, thus they are sometimes challenging to model a complicated non-linear variation in tourist volume data [18,22].

AI models are the computing techniques that simulate the intelligence process of the human brain [30,31,32], and the models such as ANN and support vector regression (SVR) have been applied to tourist volume forecasting. Many empirical results show that AI models show better performance compared to traditional time-series and econometrics models [33,34,35]; however, there is still evidence demonstrating that ARIMA families can outperform some AI techniques (i.e., ANN) in many cases [36,37]. Another AI method, SVR, is the regression version of support vector machine. Its effectiveness in capturing the non-linear trend of time-series data has been proved by many studies [8,38]. However, SVR is prone to overfitting and unqualified for training based on a large dataset. Moreover, the method pertains to shallow machine-learning techniques, which depend on manual operations for learning rather than automatically discovering the regularity in information from data [39,40].

Owing to the rapid development of AI, several prior research about tourist volume forecasts have been done based on deep learning approaches [9,11,13,41]. The deep learning models contain deep neural networks, which have a more complicated structure with more hidden layers than conventional (shallow) neural networks. Therefore, they can simulate complex non-linear variations and extract useful features from data with minimal manual intervention while learning.

Particularly, deep learning techniques that are mainly used for sequence data (e.g., tourist volume) forecasting such as recurrent neural network (RNN) and long short-term memory neural network (LSTM NN) have been proposed for the field of tourist volume forecasting. For example, Law et al. (2019) predicted monthly tourist arrival volume in Macau based on LSTM and attention mechanism [13]. Bi, Liu, and Li (2020) applied LSTM to forecast the daily tourist volume of Jiuzhaigou scenic spot and Huangshan Mountain Area [11]. Kulshrestha, Krishnaswamy, and Sharma (2020) used bidirectional LSTM to forecast the quarterly volume of tourist arrival in Singapore [41]. Deep learning methods always show a better forecast performance than the traditional modeling approaches, and owing to its effectiveness and adaptability, the usage of deep learning has become a new trend in tourist volume forecasting.

2.2. Data for Tourist Volume Forecasting

As this study is an AI-based tourist volume forecasting, this subsection will focus on reviewing the studies about forecasting data from an AI modeling perspective. Most AI-based tourist volume forecasts are data-driven approaches, and temporal characteristics of the forecasts (e.g., temporal resolution of the forecasting) are mostly dictated by data. The data for AI-based tourist volume forecasting can be divided into tourist volume data and auxiliary data, referred to as the time-series data of tourist volume and the data that corresponds with correlated variables to tourist volume (e.g., web search data), respectively.

The most common tourist volume data used by the related studies are from official statistics that are recorded and released by the related statistics departments of scenic spot or the government, which is usually collected from the official website [9,11], official publications [12] or directly provided by the official agency [10,42]. The frequency of tourist volume data from official statistics mainly ranges from daily to monthly. Majority of the studies use monthly data, and generally, the corresponding forecasting horizon is equal to or longer than a year. Hitherto, a day is recognized as the highest temporal resolution for the official statistics data, and a month is considered the minimum forecasting horizon based on such data; web search data are the most commonly used auxiliary data provided by Baidu index or Google trends. The correlation between travel information search and tourist behavior has been mentioned in many studies [43,44], and incorporating such data into training dataset helps enhance the forecasting performance [2,7,16]. However, daily is also the upper frequency limit for web search data. Additionally, whether there is a correlation between web search and ultra-short-term (e.g., less than a day) tourist flow has not been verified. Therefore, at least to date, the conventional dataset combined with the official statistics and web search data is incapable of conducting the high-temporal-resolution forecast for tourist volume.

In terms of the application of high-frequency forecasting, two types of data with high temporal granularity have been applied in related fields such as crowd forecasting and vehicle volume forecasting, which are mobile location data and traffic sensor data. The application of such data can serve as a reference to high-frequency tourism forecasting.

Mobile location data (or mobile signaling data) becomes available owing to the GNSS positioning techniques and wide usage of mobile Internet. The data have recently been applied to many short-term crowd forecasts. For example, Wang et al. (2019) forecasted crowd flow based on cellular data with a temporal resolution of 30 min [45]; Cecaj, Lippi, Mamer and Zambonelli (2020) predicted crowd distribution using aggregated mobile data with a frequency of 15 min [46]; Fan, Song, Shibasaki, and Adachi (2015) used an anonymous GPS data with time granularity of 5 min to forecast crowd movement [47]. Traffic sensor data are mostly applied to vehicle volume forecasting, and according to the related studies, such data can achieve high temporal resolution of 5 min [48,49]. However, the data collection strongly relies on tremendous sensors such as inductive-loop traffic detector, and similar sensors for crowd monitoring or pedestrian counting has not been pervasively equipped in tourist attractions at present. However, with the rapid development of technologies regarding crowd monitoring and visual crowd analysis, such data availability in near future with high reliability and frequency for tourism forecasting is expected. Then, many spatiotemporal forecasting techniques such as Conv-LSTM [50], PredRNN [51], and traffic-flow forecasting approaches [52,53] can be introduced to promote the development of the tourist flow forecasting field.

3. Materials and Methods

3.1. Control Mechanism for Forecasting Scope Based on Location-Based Services

Location-based services (LBS) data are a type of mobile location data that records real-time geographical location of mobile Internet users through wireless communication network or GNSS. In this process, each smartphone device is considered as a mobile sensor reflecting its user’s location information. This allows the collection of a large amount of individual location data in real-time. The advantages of LBS data are, it is timely, objective, massive in volume, and more importantly, offers a finer spatial and temporal resolution compared to the traditional data such as the official statistics data. According to the studies that adopted such data to conduct high-frequency crowd forecasting, temporal resolutions of the data can reach minute level [46,47].

Based on the aforementioned characteristics of LBS data, this study proposed a control mechanism to manipulate the spatial scope of the forecasts. Specifically, after delimiting the forecasting sub-area in park, the mechanism is able to produce high-frequency tourist-volume sequence of the designated area, thereby supporting the subsequent forecasts. The control mechanism illustration is shown in Figure 1. The mechanism consists of the following steps. First, at each time note (t₁, t₂, t₃…), the mechanism iterates over all the LBS data, based on which it creates LBS-based matrices and then aligns these matrices to the rasterized park area. After delimiting the spatial scope for forecasts in the park, the mechanism extracts corresponding data from the matrices and converts it to tourist volumes. This process is repeated at each time note to produce a time series sequence of tourist volume for the designated area, which is utilized as training data for deep-learning forecasting techniques. This approach can be regarded as a transfer tool that effectively connects LBS data to forecasting models to achieve maneuverable scope control for forecasts, which is able to contrapuntally assist crowd control of the local area in park.

3.2. The Proposed Deep Learning Model for High-Frequency Forecasting

3.2.1. The Neural Network Architecture

The proposed network, referred to as attention-based deep bidirectional gated recurrent unit (Att-DBGRU), integrates multiple deep-learning components to form a deeper and more effective model for the high-frequency tourist-volume forecasts. The two main components adopted in the network were deep bidirectional GRU (DBGRU) and attention mechanism. The former is composed of bidirectional GRUs (BGRUs), which can learn the tourist volume sequence data from both past and future directions to better model the sequential dependency, and its superiority compared to unidirectional RNN based models have been tested by related studies [41]. Thus, we used multiple BGRUs and stacked them to a deeper DBGRU network to further improve its accuracy and robustness. The latter enables the network to differentiate feature importance of high-frequency tourist-volume sequence and enhance the performance on capturing its long-range dependencies.

More concretely, Att-DBGRU is composed of four layer types: BGRU, attention, dropout, and fully connected layers (Figure 2). BGRUs can fully learn features from high-temporal-resolution tourist volume. The information is received from the previous BGRU layer to yield outputs, which are then transmitted to the next BGRU layer as inputs. The process is repeated in the network to fully extract valuable information from the tourist volume. Moreover, the attention layer focuses on important information in the tourist-volume sequence; dropout layers are added to the network to alleviate over-fitting, which is always considered a serious problem in machine learning and deep learning [54]. The layers randomly discarded a specific portion of the neurons in a deep learning layer to regularize and enhance the robustness of the deep network; a fully connected layer is disposed to reshape the outputting information and accomplish the modeling process of the deep network.

According to the characteristics of high-temporal-resolution forecast for tourist volume, the proposed forecasting models are many-to-one models, which use multiple inputs to yield a single output. The input features contained historical tourist volume and time-series variables correlated to tourist attraction visits, such as weather and vacation variables, and the predicted tourist volume was the only output feature.

3.2.2. Deep Bidirectional Gated Recurrent Units

The forecasting model involved multiple bidirectional RNNs to manage high-temporal-resolution sequence of tourist volume, the network proposed by Schuster and Paliwal (1997) [55], which had been applied to many time-series forecasts and always yielded outstanding performance [56,57]. With the same or similar setup parameters, a bidirectional network achieved more desired results than a unidirectional network in many fields [58].

Different from the typical unidirectional RNNs, bidirectional network consists of two separated hidden RNN layers (the forward and backward layer) in its architecture, and each of these layers is connected to an input and an output layer, respectively. The bidirectional architecture allows its network to learn the tourist volume sequence from both past and future directions. A forward and a backward layer in the network read input sequence x$\left(x_1,x_2,\dots x_t\dots ,x_{n-1},x_n\right)$ from two opposite directions, in which $x^{forward}=\left(x_1,x_2,\dots x_t\dots ,x_{n-1},x_n\right)$ and $x^{backward}=\left(x_{\mathrm{n}},x_{\mathrm{n}-1},\dots x_t\dots ,x_2,x_1\right)$, then obtain a forward hidden state ${\overrightarrow{h}}_t\left({\overrightarrow{h}}_1,{\overrightarrow{h}}_2,\dots ,{\overrightarrow{h}}_{n-1},{\overrightarrow{h}}_n\right)$ and a backward hidden state ${\overleftarrow{h}}_t({\overleftarrow{h}}_1,{\overleftarrow{h}}_2,\dots ,{\overleftarrow{h}}_{n-1},{\overleftarrow{h}}_n)$, respectively (see Equations (1) and (2)). Subsequently, forward and backward sequences are concentrated and generate the final output sequence y $\left(y_1,y_2,\dots y_t\dots ,y_{n-1},y_n\right)$, which is calculated using Equation (3).

$${\overrightarrow{h}}_t=f\left(W_{x\overrightarrow{h}}·x_t+W_{\overrightarrow{h}\overrightarrow{h}}·{\overrightarrow{h}}_{t-1}+b_{\overrightarrow{h}}\right)$$

(1)

$${\overleftarrow{h}}_t=f\left(W_{x\overleftarrow{h}}·x_t+W_{\overleftarrow{h}\overleftarrow{h}}·{\overleftarrow{h}}_{t+1}+b_{\overleftarrow{h}}\right)$$

(2)

$$y_t=W_{y\overrightarrow{h}}·{\overrightarrow{h}}_t+W_{y\overleftarrow{h}}·{\overleftarrow{h}}_t+b_y$$

(3)

where ${\overrightarrow{h}}_t, {\overleftarrow{h}}_t$, are forward and backward propagation, respectively; f ($·$) denotes a non-linear activation function (i.e., sigmoid function); $W_{x\overrightarrow{h}}$, $W_{\overrightarrow{h}\overrightarrow{h}}$, $W_{x\overleftarrow{h}}$, $W_{\overleftarrow{h}\overleftarrow{h}}$, $W_{y\overrightarrow{h}}$, and $W_{y\overleftarrow{h}}$ represent the corresponding weight coefficients and $b_{\overrightarrow{h}}$, $b_{\overleftarrow{h}}$, and $b_y$ represent corresponding bias vectors.

GRU cells were adopted to add to the aforementioned bidirectional network, which is known as an enhanced version of RNN cell and a variant of LSTM cell. Even GRU was rarely applied to tourist volume prediction, it achieved the same desired forecast effect as LSTM in many other time-series forecasts [59]. GRU simplifies the gating mechanism from three LSTM gates—forget, input, and output—to two gates—reset gate and update gate—to reduce the calculation cost. A typical GRU cell has two gates: reset gate and update gate. Reset gate determines what information the current step can access from $h_{t-1}$ and $x_t$, which is calculated using Equation (4). Subsequently, a candidate vector, ${\tilde{h}}_t$, is produced by tanh function, during which outputs of reset gate, $r_t$, only influences $h_{t-1}$ (see Equation (5)). Furthermore, update gate is used to control the influence of the previous state, $h_{t-1}$, and candidate vector, ${\tilde{h}}_t$, on state vector, $h_t$, calculated using Equations (6) and (7).

$$r_t=\sigma \left(W_r·\left[h_{t-1},x_t\right]+b_r\right)$$

(4)

$${\tilde{h}}_t=tanh\left(W_h·\left[r_t ⨀ h_{t-1},x_t\right]+b_h\right)$$

(5)

$$u_t=\sigma \left(W_u·\left[h_{t-1},x_t\right]+b_u\right)$$

(6)

$$h_t=\left(1-u_t\right) ⨀ h_{t-1}+u_t ⨀ {\tilde{h}}_t$$

(7)

where $r_t$ and $u_t$ represent the output of reset gate and update gate, respectively; $h_{t-1}$ and $h_t$ represent previous cell state and current cell state, namely; ${\tilde{h}}_t$ denotes a candidate activation vector; $W_r$, $W_h$, and $W_u$ represent weight matrixes, and $b_r$, $b_h$, and $b_u$ represent bias vectors; ⨀ denotes the Hadamard product.

3.2.3. Attention Mechanism in the Network

The attention mechanism mimics the way the human brain processes information and concentrates only on the important parts of the information while ignoring the irrelevant parts. We utilized the mechanism to highlight the key parts of time-series sequence of high-frequency tourist volume. Specifically, our attention mechanism was added after the DBGRU layers to assign probability weights to the output vectors from the last DBGRU layer (Figure 2). Thus, the neural network can focus more on the important content of the vector to yield highly effective forecasts. This study used Luong’s multiplicative attention to construct our custom attention layer, and the calculating processes are listed as follows (see Equations (8)–(11)).

$$a_t=\frac{\exp \left(score\left(h_t, {\overline{h}}_s\right)\right)}{{{\displaystyle \sum }}_{s^{\prime }=1}\exp \left(score\left(h_t, {\overline{h}}_{s\prime }\right)\right)}$$

(8)

$$score\left(h_t, {\overline{h}}_s\right)=h_t{}^T{W}_a{\overline{h}}_s$$

(9)

$$c_t={{\displaystyle \sum }}_{s=1}^T{a}_t{\overline{h}}_s$$

(10)

$${\tilde{h}}_{att}= \tan \mathrm{h}\left(W_c\left[c_t;h_t\right]\right)$$

(11)

where $h_t$ and ${\overline{h}}_s$ denote current hidden state and source hidden states, respectively; $a_t$ represents attention weight; $c_t$ represents context vector; ${\tilde{h}}_{att}$ denotes attention vector.

4. Experiment

4.1. Overview

The objective of the experiment is to verify the effectiveness of the proposed method. Specifically, we delimit three forecasting areas and produce their high-frequency tourist volumes based on the proposed scope control mechanism. We then use Att-DBGRU model to forecast a 3-day tourist volume of the designated areas in a temporal resolution of 30 min, as we believe this forecasting provides sufficient information about the short-term high-frequency variation of the tourist volumes for the park managers to conduct crowd control decision-making. Finally, the performance of the proposed method is evaluated based on multiple measurements, which is then compared with other forecasting techniques. More specific procedures are illustrated in Figure 3.

4.2. Study Area

Taiyangdao Scenic Spot is located in Harbin—the capital of Heilongjiang province, China. It has been given a 5A rating—the highest level in China’s tourist attraction rating system according to the Ministry of Culture and Tourism of the People’s Republic of China (https://zwfw.mct.gov.cn/scenicspot/ (accessed on 10 May 2021)). This study focuses on one of its well-known sub-area, the Taiyangdao Park. The three forecasting areas are the entire Taiyangdao Park and its two sub-regions (Area 1 and Area 2). Area 1 and Area 2 are crowd-prone areas that are always highly congested during the peak hours of the day, therefore short-term forecasts of the areas help practically in making crowd control decisions.

The spatial layout of Taiyangdao Park is presented in Figure 4. The area in yellow represents the entrance to the park (Area 1), which has four entrance gates (represented by the red dots). Furthermore, the area in red represents the Taiyang lake area (Area 2), which is also known as the central area. It contains the main recreational facilities such as the Taiyang lake (the largest lake in the park), amusement park, and central square. Figure 4 also contains aerial photographs which provide more information on these areas.

4.3. Data Preparation and Model Setting

China’s mobile Internet users have reached 1.319 billion by the end of 2019 according to the 2020 China Mobile Internet Development Report. With the widespread use of mobile Internet in China and advanced mobile positioning techniques, mobile location data such as LBS data can reflect real-time variations in tourist volume in Chinese tourist attractions, and this becomes highly valuable in applying relevant time-series or spatio-temporal forecasting. This study procured LBS data from the Tencent Company—one of the well-known Internet service companies with the largest user base in China. Tencent LBS provides various big data-related products, such as the platform of Tencent location big data (https://heat.qq.com/bigdata/index.html (accessed on 3 March 2021)). The regional heatmap (https://heat.qq.com/heatmap.php (accessed on 3 March 2021)) is one of the services offered by Tencent location big data, allowing the collection of high-resolution location data of the tourists in Taiyangdao Park. We collected data in a 30-day interval (from 21 September 2020 at 5:00 AM to 21 October 2020 at 4:30 AM) to set up a dataset for each of the study area (i.e., Taiyangdao park, Area 1 and Area 2). Each dataset had a sample size of 1440 and a 30-min temporal resolution.

Many studies note that variables related to weather and vacation influence the travel behavior [60,61], and adding these features into inputs of model training can enhance the accuracy of tourist volume prediction [62,63]. Hence, we considered these variables as parts of our input features. First, we collected the weather data from Worldweatheronline (https://www.worldweatheronline.com/ (accessed on 11 March 2021)); the data show that there is no very extreme weather condition in our study area during the period, and primary weather types are sunny, cloudy, and rainy. We converted the nominal weather variable to dummy variables by binarization. As tourist activities in Taiyangdao park mainly occur outdoors and can be more impacted by rainy weather, we used different dummy variables to represent rainy and others, where 0 represents rainy and 1 represents sunny and cloudy. Moreover, the nominal vacation variable were also converted to a binary value where 1 represented vacation days and 0 represented the rest. In addition, we split the total samples to set up the training and test dataset. The former was used for model training, and the latter was prepared to evaluate forecast performance. The data for the last three days was used to set up the test dataset with a size of 144, and the previous data were used for model training.

Our dataset contained multi-dimensional variables, including tourist volume, weather, and vacation. Therefore, the data was normalized before being sent into the deep learning models. We used the min-max normalization to normalize the original data, transforming the value of features to dimensionless values within [0,1] range to reduce the sensitivity of the model to the scale of features. A de-normalization operation was also conducted after the training to inverse transform all the normalized data to its original scale.

With regards to the forecasting models, this study along with verifying the effectiveness of the proposed Att-DBGRU validates the superiority of Att-DBGRU compared to other commonly used deep-learning models in this high-frequency forecasting, which is also considered an objective of this study. The benchmark models adopted are divided into two general-RNN based models, (general-RNN model) based on typical RNN network, and bidirectional-RNN based models (bi-RNN models) based on bidirectional RNN networks. The general-RNN includes some widely used model in tourist volume forecasting and tourist demand forecasting field (such as LSTM), the bi-RNN models we adopted are DBGRU and DBLSTM, the latter were also used in other studies of tourist demand forecasting [41].

For all the general- and bi-RNN models, the activation function is used a rectified linear unit (ReLU)—this function can reduce the probability of vanishing gradient and generate sparse networks to prevent overfitting [64]. Additionally, we employed a regularization for each hidden layer to reduce the overfitting based on a dropout method. The dropout rate was set to 0.1, meaning that 10% of the neurons were randomly discarded. Furthermore, we adopted adaptive moment estimation (Adam) as the optimization method, which has been widely applied in many related studies [65,66,67]. The predicting experiments and the training were executed using Python 3.9, Anaconda, TensorFlow, and PyTorch.

4.4. Performance Evaluation

This study adopted the following three measures to evaluate the performance of the forecasting models: fitting methods, error measures, and Diebold–Mariano tests. The fitting methods including the fitting curves and scatters can visualize the forecasting results and real observations of tourist volume, which directly show the fitting effects of different periods; the error measures include the root means square error (RMSE), the mean absolute error (MAE), and the coefficient of determination (R²). Their mathematical formulations are presented in Equations (12)–(14).

$$\mathrm{RMSE}=\sqrt{\frac{1}{N}·{{\displaystyle \sum }}_{i=1}^N{\left(y_i-{\widehat{y}}_i\right)}^2}$$

(12)

$$\mathrm{MAE}=\frac{1}{N}·{{\displaystyle \sum }}_{i=1}^N|y_i-{\widehat{y}}_i|$$

(13)

$$R^2=\frac{{{\displaystyle \sum }}_{i=1}^N\left(y_i-\overline{y}\right)\left({\widehat{y}}_i-\overline{Y}\right)}{\sqrt{{{\displaystyle \sum }}_{i=1}^N{\left(y_i-\overline{y}\right)}^2}\sqrt{{{\displaystyle \sum }}_{i=1}^N{\left({\widehat{y}}_i-\overline{Y}\right)}^2}}$$

(14)

where N represents the value of the sample size; $y_i$ and ${\widehat{y}}_i$ denote observed tourist volume and forecasted tourist volume, respectively;$\overline{y}$ and $\overline{Y}$ represent the mean of $y_i$ and ${\widehat{y}}_i$, respectively. The smaller the values of RMSE and MAE are, the better the performance of the forecasting model is. The closer the value of R² to 1 is, the more accurate the forecast.

Diebold–Mariano test (D-M test) is adopted to validate if the performances of predictions are significantly different, which can also imply the difference of model capacity in such high-frequency forecasting for the park and its sub areas.

5. Results and Discussion

5.1. Fitting Effects of the Forecasting Techniques

The empirical study was focused on Taiyangdao Park and its two designated sub-areas using the proposed Att-DBGRU method and other five deep learning methods—DBGRU, DBLSTM, GRU, LSTM, and RNN. Figure 5 demonstrates the fitting effects for a visual comparison between the forecasts based on the six models and the real observation. The figure illustrates the actual and forecasted tourist volume of the entire Taiyangdao Park on the last three days of the entire 30-day time horizon with a temporal resolution of 30 min, where the black line represents the actual 144 observations of tourist volume.

In the forecast results for the entire Taiyangdao Park, Att-DBGRU generated ideal fitting effects. Its forecasting curve fit the real observations well even in peaks and valleys where the curves of many other methods deviated from the actual values to some extent. The scatter of the Att-DBGRU also illustrates that its forecasts fit the real observation well in both high- and low-volume intervals. The good fitting effects indicate the superiority of Att-DBGRU to the high-frequency forecast of tourist volume of the park. Many other deep-learning methods also display the desired effects on fitting the actual observation during the period, particularly DBGRU and DBLSTM, which yielded well-fitted values in the low-volume interval; however, the values deviated slightly from the real observation line in the high-volume interval (800–1500). Generally, the bi-RNN models perform more accurate forecasts than the general-RNN models. The latter generated well-fitted values in the 200–600 interval but did not fit well in the high-volume interval and extremely low-volume (0–100) interval according to the scatter.

With regards to the forecasts at the sub areas in the park, the results for the two sub areas are illustrated in Figure 6. The results are comparable to the previous results, showing that Att-DBGRU makes accurate forecasts for both the sub areas, although its scatter in area 1 deviated slightly from the actual value in approximately 800–1000 interval, but the general results are more well-fitted than the result of other methods, particularly in the high-volume interval. This indicated the obvious advantages of Att-DBGRU in peak period forecasting of the tourist volumes compared to the other deep learning models. In terms of the Bi-RNN models, although the curves and scatters display relatively significant deviations than Att-DBGRU, the overall performances are still satisfying in the sub-areas forecasting tests and more precise than the performance generated by the general-RNN models. Particularly, in the area 2, the forecasts by bi-RNN models slightly deviated from the actual value, but the forecasts by some general-RNN models such as GRU and RNN shows more significant deviations compared to the bi-RNN models.

5.2. Results of the Performance Evaluations

Three performance measures—RMSE, MAE, and R2—were used to evaluate the performance of the forecasting models.

Table 1. Results of the performance evaluations (Taiyangdao Park).

Model		RMSE	MAE	R2
Name	Category of Network
RNN	Unidirectional	146.138	122.510	0.655
LSTM	Unidirectional	118.329	95.419	0.816
GRU	Unidirectional	118.395	93.851	0.839
DBLSTM	Bidirectional (Stacked)	102.478	72.101	0.896
DBGRU	Bidirectional (Stacked)	79.326	57.003	0.939
Att-DBGRU	Bidirectional (with attention)	62.289	44.482	0.970

shows the results of these measures for the forecasts of Taiyangdao park. According to the table, Att-DBGRU made the best forecast with the lowest RMSE and MAE scores (62.289 and 44.482, respectively) with a highest R2 score of 0.970. This not only indicated the effectiveness of Att-DBGRU in such high-frequency forecasts, but it also outperformed the bi-RNN and general-RNN methods. Additionally, bi-RNN models outperformed general-RNN models, particularly DBGRU, which also achieved desired outcomes; although its result (the RMSE, MAR, and R2 score are 79.326, 57.003, and 0.939, respectively) is less accurate than Att-DBGRU, it is distinctly more accurate compared to DBLSTM and the general-RNN models. In terms of the general-RNN models, the RMSE, MAE, and R2 scores of GRU and LSTM are approximately equivalent to each other, and both demonstrated much higher performance accuracy compared to that of RNN.

The performances for Area 1 and Area 2 are shown in

Table 2. Results of the performance evaluations (Area 1 and Area 2).

	Model		RMSE	MAE	R2
	Name	Category of Network
Area 1	RNN	Unidirectional	94.887	71.858	0.714
	LSTM	Unidirectional	87.392	64.002	0.780
	GRU	Unidirectional	79.978	57.057	0.808
	DBLSTM	Bidirectional (Stacked)	72.413	49.038	0.897
	DBGRU	Bidirectional (Stacked)	66.724	45.475	0.908
	Att-DBGRU	Bidirectional (with attention)	45.583	34.008	0.959
Area 2	RNN	Unidirectional	39.929	30.804	0.824
	LSTM	Unidirectional	37.071	33.436	0.862
	GRU	Unidirectional	45.108	38.811	0.859
	DBLSTM	Bidirectional (Stacked)	30.440	23.608	0.929
	DBGRU	Bidirectional (Stacked)	28.881	19.354	0.926
	Att-DBGRU	Bidirectional (with attention)	16.622	9.761	0.981

, which shows that Att-DBGRU also achieved ideal outcomes in forecasting the sub areas in the park, with the highest R2 score of 0.959 and 0.981, respectively. The lowest RMSE and MAE scores of Att-DBGRU also demonstrated it had the best result among all the models in both of the sub areas. This illustrated the desired forecasting capacity of the proposed method for such high-frequency forecasts of tourist volumes for the designated sub-areas in the park. Similar to the results of the evaluation in the entire Taiyangdao Park, the bi-RNN models provide better predicting performance than the general-RNN models based on the three evaluation measures, and the performances of the two bi-RNN models are approximately comparable with each other. In addition, GRU and LSTM have minimal forecasting errors among the general-RNN models in Area 1 and Area 2, respectively, and both of them outperform RNN.

5.3. Comparison of Forecasting Methods with Diebold–Mariano Test

This study used D-M test to validate each model with every other model in turn, to compare if one significantly outperforms the other in the high-frequency forecasting. The utilized D-M test is based on MAE and MSE error statistic, respectively (

Table 3. Results of Diebold–Mariano test.

Area	Model	DMCriteria	Benchmark Model
			DBGRU	DBLSTM	GRU	LSTM	RNN
	Att-DBGRU	(MAE)	−1.114	−1.881 *	−3.194 **	−3.628 **	−5.073 **
		(MSE)	−1.118	−1.878 *	−2.560 **	−3.087 **	−3.400 **
	DBGRU	(MAE)		−2.672 **	−7.297 **	−4.462 **	−7.585 **
		(MSE)		−2.962 **	−4.060 **	−4.452 **	−4.309 **
Taiyangdao	DBLSTM	(MAE)			−3.867 **	−1.775 *	−5.513 **
		(MSE)			−5.031 **	−1.825 *	−4.177 **
	GRU	(MAE)				−0.165	−3.756 **
		(MSE)				0.008	−2.871 **
	LSTM	(MAE)					−3.067 **
		(MSE)					−2.610 **
	Att-DBGRU	(MAE)	−2.709 **	−2.724 **	−3.235 **	−4.200 **	−4.033 **
		(MSE)	−2.936 **	−2.444 **	−2.976 **	−3.145 **	−2.966 **
	DBGRU	(MAE)		−1.320	−1.663 *	−3.720 **	−3.831 **
		(MSE)		−1.227	−1.864 *	−2.819 **	−2.613 **
Area 1	DBLSTM	(MAE)			−1.194	−3.319 **	−3.722 **
		(MSE)			−1.405	−3.108 **	−2.854 **
	GRU	(MAE)				−1.642	−1.893*
		(MSE)				−1.780 *	−2.044 **
	LSTM	(MAE)					−1.685 *
		(MSE)					−1.678 *
	Att-DBGRU	(MAE)	−3.565 **	−8.862 **	−6.526 **	−5.577 **	−7.648 **
		(MSE)	−2.267 **	−3.501 **	−4.886 **	−5.610 **	−3.366 **
	DBGRU	(MAE)		−2.065 **	−2.972 **	−2.202 **	−3.462 **
		(MSE)		−0.525 *	−2.375 **	−1.441	−2.516 **
Area 2	DBLSTM	(MAE)			2.610 **	−1.910 *	−2.770 **
		(MSE)			−2.225 **	−1.345	−2.623 **
	GRU	(MAE)				1.221	1.352
		(MSE)				1.532	0.686
	LSTM	(MAE)					0.523
		(MSE)					−0.436

* and ** denotes the significance level of 10% and 5%, respectively.

). A negative DM value denotes the corresponding model tends to outperform the benchmark model, and vice versa if the DM value is positive.

According to the results of the D-M tests, the proposed Att-DBGRU outperforms all the other models at 5% significance level in the two sub areas, which indicates its significant superiority in high-frequency forecast of the sub areas compared to other deep-learning methods. However, in the entire Taiyangdao Park, although Att-DBGRU significantly outperforms DBLSTM and the general-RNN models at the 10% or 5% significance level, the superiority of the method to DBGRU is not significant. This implies DBGRU model also has satisfactory capacity in such forecasting and can be considered an alternative to Att-DBGRU in some cases.

In terms of comparison of the other five models, DBGRU outperforms almost all the general-models at 10% or 5% significance level, and DBLSTM also performs significantly accurate forecasts compared to the general-RNN models in most cases, the only exception occurs in Area 1, both the MAD- and MSE-based D-M tests show that the performance difference between DBLSTM and GRU is not significant, which also implies that forecasting model can be differentiated depending on the region that is being forecasted. Generally, bi-RNN models provide better performance than general-RNN models, which demonstrates the former have significant superiority in such high-frequency forecasts of park visitation compared to the frequently used general-RNN models (such as LSTM) by the related works.

5.4. Discussion

Most related studies focus on improving the forecasting model. Earlier, they adopt linear methods (i.e., ARIMA and its variants) [20,29] and econometrics methods [12], and then shallow learning methods [8,38]. Recently, deep learning models have achieved outstanding performances in many tourist volume forecasting [9,11,13]. The application of more advanced model largely enhances the forecasting accuracy, which provides vast contribution to the field. However, improving the controllability and flexibility of the forecasting method is of equal importance, which is neglected all along. Particularly, the control of the forecasting scope can make the forecasting more focalized, and can provide the forecasting information of tourist volume at a finer sub-site level.

An example of how the results of spatially controllable forecasts can contribute to the actual crowd management is provided to emphasize its practicality. Based on the forecasting information at sub-site scale, we can show density comparisons between the entire site and its two sub-areas (Figure 7). According to the hourly illustration of forecasted crowd density, we can observe that the two sub-areas experience a much higher density than the entire site during a specific period of the day (Area 1 at 11:00, 12:00, and 15:00 and Area 2 at 11:00–15:00). This information regarding high density period can contribute to the short-term decision in preventing crowding at local scale, and this information cannot be reflected by forecast of the entire site. The highly specific information can support the discovery of the disparity in crowd density between the sub-regions of the park within a specific period and advise park management to shift their focus from one sub-region to another.

With regards to the forecasting model, this study also proposed a novel deep-learning model (Att-DBGRU) for the high-frequency forecasting, and compared its effectiveness with other deep learning models used by related works. According to [9,11,13,68,69], deep learning techniques always outperform the shallow learning methods and linear methods in the forecasting. Therefore, we adopted only the deep learning models as our benchmark models. Notably, the deep learning algorithms used in this study were categorized into general-RNN and bi-RNN models, to allow a comparison between different models. These models are commonly used by other tourist volume forecasts, such as LSTM [9,11] and BLSTM [41]. The results of this study validated the significant forecasting capacity and superiority of the proposed Att-DBGRU in high-frequency forecasts, and demonstrated that such method could be a new forecasting approach of tourist volume. Furthermore, the findings provide more evidence in the differences between widely used deep learning methods in such high-frequency forecasting performance. Generally, bi-RNN models outperformed the general-RNN models in forecasting, and such result is consistent with the works from Kulshrestha, Krishnaswamy, and Sharma (2020) [41], who conducted a long-term tourism demand forecasting based on bidirectional LSTM.

6. Conclusions

This study proposed a novel method based on a control mechanism and a forecasting model named Att-DBGRU to achieve spatially controllable high-frequency forecasts of park visitation. To validate the effectiveness of the method in high-frequency forecasting and test the superiority of Att-DBGRU compared to other widely used deep learning methods, we experimentally evaluated the method considering Taiyangdao park and its two subareas. Moreover, we adopted fitting methods, error measures, and D-M tests to evaluate the forecasting performances.

The experimental results reveal that the proposed method could provide outstanding performance in high-frequency forecasts of tourist volume for the three designated areas, indicating that this method is suitable for implementation for the forecast of the park and its sub-areas. Furthermore, the proposed Att-DBGRU model significantly outperformed other commonly used deep learning methods in tourist volume forecasting. Moreover, the findings also involved more evidence about the forecasting accuracy of other deep-learning techniques, for example, bi-RNN models outperformed unidirectional general-RNN models in such high-frequency forecasting. Particularly DBGRU, which also obtained satisfying outcomes in the forecasting experiments, and can be an alternative model for such high-temporal-resolution forecasts. In addition, GRU demonstrated the most accurate prediction among those of the general-RNN models and outperformed LSTM, the most commonly used deep learning model in the tourist volume forecasting field.

The study’s contributions to the field of tourist volume forecasting are: First, it achieved the forecast for designated sub-areas in the park, and the design of the control mechanism of the proposed method provided a more controllable and flexible forecast, actualizing the forecast at a finer spatial scale. Second, the study performed high-frequency forecasts of tourist volume, with a temporal resolution up to a minute-level. Third, an advanced deep learn model (Att-DBGRU) was proposed, which significantly outperformed the widely used bi-RNN and general-RNN models in high-frequency forecasting.

The practical implications of this study are as follows: First, the high-frequency forecast predictions can contribute to multiple short-term decision-making related to tourist attraction management, such as intraday crowd control, staffing arrangement, and accident prevention. Second, regarding the spatial scale of the forecast, forecasting only for an entire tourist attraction provides limited support in operational management. Therefore, the proposed method further focused on the sub-areas and their forecasting information, which is of practical use in local crowd control.